Research On CNN Classification Method For Diabetic Retinopathy

Retinopathy(Diabetic Retinopathy,DR)is a common complication of diabetes.Current clinical diagnosis usually divides DR into 5 grades,which is determined by an ophthalmologist by reading the abnormal lesion on the patient’s retina image.This method is time-consuming and labor-intensive and requires high clinical experience from ophthalmologists,which may result in patients not receiving timely treatment.Therefore,experts from all walks of life are constantly seeking more efficient diagnostic methods.With the development of image processing technology and machine learning,computer-aided diagnosis and treatment based on artificial intelligence has been widely used in the processing of medical images.Designing a complete image classification system and using computer-aided screening of a large number of diabetic retina images can greatly help ophthalmologists to make fast and effective diagnosis.In this paper,based on the existing model,an improved convolutional neural network(CNN)-Supplement Net is designed.The comparative analysis proves the effectiveness of the model.The following is the specific research content:(1)Systematically summarized the concepts and basic principles of CNN and transfer learning.On the basis of fully understanding the current research status at home and abroad,the system summarizes the concepts and basic principles of each layer in CNN,introduces the basic principles of transfer learning and comparatively analyzes the advantages and disadvantages of existing models such as Le Net-5,Alex Net,VGGNet and Goog Le Net.The image classification model was established with the VGG-16 network architecture as the core.(2)An image classification model of diabetic retinopathy based on convolutional neural network was established.In view of the problem that the accuracy of the existing model needs to be improved,this paper draws on the existing CNN model idea and designs a new CNN model-Supplement Net.The shallow structure of the model inherits the structural parameters of VGG-16 by means of transfer learning,the deep structure adopts the training method from zero to better learn the sample features,and then adjust the parameters.Based on the original depth model,the Supplement Net model improves the activation function in the convolutional layer to ensure the nonlinearity of the network and batch-regularize the image data after the corresponding convolution module to improve the generalization performance of the model.Finally,a comparative experiment is designed for different models and training methods.The results show that the network performance of the Supplement Net model is better than the traditional model,and the two classification and five classification problems have obtained a higher accuracy.(3)An image classification system for diabetic retinopathy was designed.In view of the inconvenience of using the network model directly and considering that the user does not understand the parameter setting when training the model and the specific use steps during classification,etc.,this article uses MATLAB visual graphics window tool to design the DR image classification system according to user needs.First,the GUI interface layout was established in combination with the use steps and requirements of the model.Secondly,the code was edited in the callback function of each component to achieve the corresponding function.Finally,a system combining model training and image classification was formed and verified.